GitHub Trending on August 15: diagram-design still leads, plus three free developer offers (2026)

A one-day star burst put one diagram toolkit far ahead of the rest: diagram-design gained 3,646 stars on GitHub's daily board. Behind it, the useful split is between auditable agent infrastructure, spec-driven coding, local model work, and tools that turn developer hardware into a smaller surface area. 1

The fastest climbs

The table is sorted by stars gained in GitHub's daily window, not by the board's display order. Total stars and primary language are the values shown on the board. The final column adds the README detail that tells you whether a repository deserves a test or only a bookmark.

Rank	Repository	Stars gained	Total stars	Language	What it is and the first useful move
1	cathrynlavery/diagram-design	+3,646 1	17,446 1	HTML	An agent skill with 27 visual types for Claude Code, Codex, and Pi. The current README adds semantic system patterns and optional accessible motion while keeping static HTML as the default. Open it if your current agent produces generic architecture boxes. 2
2	semantica-agi/semantica	+1,181 1	7,580 1	Python	A self-hostable context and knowledge-graph layer for agents, with decision records, provenance, conflict detection, and explainable reasoning. The README's quick start is `pip install semantica`; treat it as infrastructure for auditable systems, not a small RAG helper. 3
3	github/spec-kit	+1,160 1	128,590 1	Python	GitHub's open-source toolkit for a spec-driven process with AI coding agents. Start here when the problem is vague requirements and rework, rather than a missing model or editor plugin. 4
4	holaboss-ai/holaOS	+769 1	7,353 1	TypeScript	A local-first workspace that runs Claude Code, Codex, or holaOS over your tools, files, and shared memory, with built-in models or your own keys. It supports macOS, Windows, and Linux; read the modified Apache 2.0 license before embedding it in a product. 5
5	cactus-compute/needle	+662 1	5,678 1	Python	A 45M-parameter tool-calling and structured-extraction model packed into a 14 MB binary. A full session is about 28 MB of RAM, and inference does not need a network after setup. Try `pip install cactus-needle` when offline or edge constraints matter more than frontier-model breadth. 6
6	lightningpixel/modly	+579 1	5,998 1	TypeScript	A local, open-source desktop app that turns a photo into a 3D mesh with AI models running on your GPU. It has installers for Windows, Linux, and Apple Silicon macOS; expect a real model-download and GPU setup step, not a browser demo. 7
7	unslothai/unsloth	+501 1	71,559 1	Python	A desktop and web UI for running, training, and deploying local LLMs, diffusion, embedding, and audio models. Its README also includes agent integrations, hardware-specific paths, and free notebooks for model training. 8
8	infiniflow/ragflow	+473 1	88,422 1	Go	An open-source RAG engine with document parsing, explainable chunking, grounded citations, agent templates, and connectors for common data sources. Use the cloud demo for a quick look; self-hosting is a separate deployment project. 9

Three worth opening first

Diagram Design is a follow-up with a reason. It was already visible in the previous daily digest, but the current README now calls out version 2.3's semantic patterns and optional motion. That is a concrete change to inspect: behavior such as a queue, policy trace, or trust boundary can be described separately from the layout, while static output remains the safe default. 2

Semantica is the serious infrastructure bet. Its README treats decisions as queryable objects and attaches W3C PROV-O provenance to facts. That makes it relevant to a regulated workflow that needs to answer "why did the agent do this?" It also means more setup and more design work than dropping a vector index into an app. 3

Needle is the cleanest constraint story. A 14 MB engine, roughly 28 MB of session memory, structured JSON output, confidence scores, and offline inference point to edge devices and private local tools. Its star count says people are looking; it does not say the model matches a frontier model, so test your own tool schemas before choosing it. 6

The rest of the board gives you different entry points: spec-kit for reducing specification drift, holaOS for one workspace around several agents, modly for local image-to-3D work, unsloth for local model training, and ragflow for document-heavy RAG. The right first click depends on the bottleneck, not the star count.

Free developer value to claim or try

1. Pullfrog: free on personal accounts and public repositories

Pullfrog's August 8 pricing update says the platform is now free on personal accounts and public repositories. The developer plan includes unlimited runs, unlimited public repositories for organizations, personal public and private repositories, PR review, issue triage, CI autofix, and built-in tools, hooks, MCPs, skills, and agents. 10

Who it fits: an individual maintainer, or an open-source organization that wants agent runs around GitHub issues and pull requests. Pullfrog's billing documentation says there is no card requirement for the free path. 11

How to start: Sign in with GitHub, install the GitHub App, choose a repository, and add the pullfrog.yml workflow. For model access, you can bring your own key or use Pullfrog Router at provider cost. The docs also name Big Pickle as a free model that needs no key. 11 12

The catch: free platform access does not make every model free. Private repositories inside GitHub organizations require Pro after the trial, and Pro is listed at $30 per month per account. Open-source maintainers can separately apply for free premium-model access by sending Pullfrog a repository link and a short description of how they would use it. 11

2. Port22: a free remote-control tier for coding agents

Port22's current Product Hunt launch page describes a phone companion for coding agents already running on a Mac. Its official site lists Claude Code, Codex, and OpenCode support, with a live transcript and phone-based approval for work that needs attention. 13 14

Free tier: the full app is free forever for one Mac and two sessions at a time. The official pricing section says there is no account or card requirement; the free tier includes replies, approvals, push notifications, and an encrypted relay when you are away from the local network. 14

Prerequisites: you need a Mac running the agent, an iPhone, and an agent CLI in a supported terminal. The site lists open-source terminals such as Ghostty, kitty, Alacritty, WezTerm, tmux, and Terminal.app, and says IDE-built-in terminals and Warp are unsupported. Get the iOS app or download the Mac agent. 14

3. Unsloth free notebooks: a zero-cost starting point for model experiments

The Unsloth README links free notebooks for Gemma 4, Qwen3.5, gpt-oss 20B, embeddings, TTS, and other workflows. It also lists a free Unsloth Studio notebook for running and training models in a web UI. 8

Action: Open the free Unsloth Studio notebook, then choose a notebook that matches the model and task you want to test. The same README documents local desktop installs for Windows, macOS, Linux, and WSL, so the notebook is a quick experiment path rather than the only way to use the project. 8

What to star, test, and claim

Star or test diagram-design if you need architecture diagrams, flows, timelines, or policy traces that look deliberate instead of auto-generated. Its new semantic-pattern and motion work is the reason to revisit it today. 2
Test needle when RAM, bandwidth, or offline operation is the hard limit. Start with pip install cactus-needle and measure your own tool-calling accuracy. 6
Inspect semantica or ragflow when your project needs a context layer: Semantica emphasizes decision provenance and reasoning; RAGFlow emphasizes document parsing, chunking, and grounded citations. 3 9
Claim the free path that matches your setup: Pullfrog for personal/public GitHub work, Port22 for one Mac and two agent sessions, and Unsloth's notebooks for a model experiment. Each has a real free boundary; read that boundary before building a workflow around it.